We had a great discussion about temperature settings during the Q&A at last night’s AI Publishing Collective meetup, and I thought it was worth unpacking a bit further here. I’ll leave the question of whether LLMs are truly deterministic to Konrad Rzezniczak and the rest of the experts!
Temperature settings control how creatively a chatbot responds. A high temperature (e.g., 1) introduces more randomness and creativity, while a low temperature (e.g,. 0) constrains randomness, resulting in more conservative and predictable output. Here’s an example using a simple prompt:
𝘗𝘭𝘦𝘢𝘴𝘦 𝘤𝘳𝘦𝘢𝘵𝘦 𝘢 𝘵𝘸𝘦𝘦𝘵 𝘰𝘧 𝘵𝘩𝘪𝘴 𝘸𝘦𝘣𝘱𝘢𝘨𝘦 𝘧𝘰𝘳 𝘮𝘦: 𝘩𝘵𝘵𝘱𝘴://𝘴𝘤𝘩𝘰𝘭𝘢𝘳𝘭𝘺𝘬𝘪𝘵𝘤𝘩𝘦𝘯.𝘴𝘴𝘱𝘯𝘦𝘵.𝘰𝘳𝘨/2025/04/29/𝘢𝘪-𝘴𝘵𝘳𝘢𝘵𝘦𝘨𝘺-𝘨𝘰𝘷𝘦𝘳𝘯𝘢𝘯𝘤𝘦-𝘢𝘯𝘥-𝘮𝘰𝘯𝘦𝘵𝘪𝘻𝘢𝘵𝘪𝘰𝘯-𝘪𝘯-𝘴𝘤𝘩𝘰𝘭𝘢𝘳𝘭𝘺-𝘱𝘶𝘣𝘭𝘪𝘴𝘩𝘪𝘯𝘨-𝘭𝘦𝘴𝘴𝘰𝘯𝘴-𝘧𝘳𝘰𝘮-𝘪𝘯𝘥𝘶𝘴𝘵𝘳𝘺-𝘧𝘳𝘰𝘯𝘵-𝘳𝘶𝘯𝘯𝘦𝘳𝘴/
Using ChatGPT-4o, which is said to have a default temperature of around 0.7, I get the following:
𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗿𝗲𝘀𝗽𝗼𝗻𝘀𝗶𝗯𝗹𝗲, 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗰 𝗔𝗜 𝗮𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗹𝗼𝗼𝗸 𝗹𝗶𝗸𝗲 𝗶𝗻 𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴? 𝗧𝗵𝗶𝘀 𝗦𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆 𝗞𝗶𝘁𝗰𝗵𝗲𝗻 𝗮𝗿𝘁𝗶𝗰𝗹𝗲 𝘀𝗵𝗮𝗿𝗲𝘀 𝗸𝗲𝘆 𝗹𝗲𝘀𝘀𝗼𝗻𝘀 𝗳𝗿𝗼𝗺 𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗹𝗲𝗮𝗱𝗲𝗿𝘀 𝗼𝗻 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲, 𝗺𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻, 𝗮𝗻𝗱 𝗰𝗿𝗼𝘀𝘀-𝘁𝗲𝗮𝗺 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻. 🔗 𝗵𝘁𝘁𝗽𝘀://𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆𝗸𝗶𝘁𝗰𝗵𝗲𝗻.𝘀𝘀𝗽𝗻𝗲𝘁.𝗼𝗿𝗴/𝟮𝟬𝟮𝟱/𝟬𝟰/𝟮𝟵/𝗮𝗶-𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆-𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲-𝗮𝗻𝗱-𝗺𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻-𝗶𝗻-𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆-𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴-𝗹𝗲𝘀𝘀𝗼𝗻𝘀-𝗳𝗿𝗼𝗺-𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝘆-𝗳𝗿𝗼𝗻𝘁-𝗿𝘂𝗻𝗻𝗲𝗿𝘀/ #𝗦𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 #𝗔𝗜𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆 #𝗚𝗲𝗻𝗔𝗜
If I specify a lower temperature in the prompt, “𝘗𝘭𝘦𝘢𝘴𝘦 𝘤𝘳𝘦𝘢𝘵𝘦 𝘢 𝘵𝘸𝘦𝘦𝘵 𝘰𝘧 𝘵𝘩𝘪𝘴 𝘸𝘦𝘣𝘱𝘢𝘨𝘦 𝘧𝘰𝘳 𝘮𝘦: 𝘩𝘵𝘵𝘱𝘴://𝘴𝘤𝘩𝘰𝘭𝘢𝘳𝘭𝘺𝘬𝘪𝘵𝘤𝘩𝘦𝘯.𝘴𝘴𝘱𝘯𝘦𝘵.𝘰𝘳𝘨/2025/04/29/𝘢𝘪-𝘴𝘵𝘳𝘢𝘵𝘦𝘨𝘺-𝘨𝘰𝘷𝘦𝘳𝘯𝘢𝘯𝘤𝘦-𝘢𝘯𝘥-𝘮𝘰𝘯𝘦𝘵𝘪𝘻𝘢𝘵𝘪𝘰𝘯-𝘪𝘯-𝘴𝘤𝘩𝘰𝘭𝘢𝘳𝘭𝘺-𝘱𝘶𝘣𝘭𝘪𝘴𝘩𝘪𝘯𝘨-𝘭𝘦𝘴𝘴𝘰𝘯𝘴-𝘧𝘳𝘰𝘮-𝘪𝘯𝘥𝘶𝘴𝘵𝘳𝘺-𝘧𝘳𝘰𝘯𝘵-𝘳𝘶𝘯𝘯𝘦𝘳𝘴/, 𝘵𝘦𝘮𝘱𝘦𝘳𝘢𝘵𝘶𝘳𝘦 = 0.1“, 𝘵𝘩𝘦 𝘳𝘦𝘴𝘶𝘭𝘵 𝘪𝘴 𝘮𝘰𝘳𝘦 𝘳𝘦𝘴𝘵𝘳𝘢𝘪𝘯𝘦𝘥:
𝗔 𝘁𝗵𝗼𝘂𝗴𝗵𝘁𝗳𝘂𝗹 𝗼𝘃𝗲𝗿𝘃𝗶𝗲𝘄 𝗼𝗳 𝗔𝗜 𝗶𝗻 𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗧𝗵𝗲 𝗦𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆 𝗞𝗶𝘁𝗰𝗵𝗲𝗻: 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆, 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲, 𝗮𝗻𝗱 𝗺𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗹𝗲𝘀𝘀𝗼𝗻𝘀 𝗳𝗿𝗼𝗺 𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗹𝗲𝗮𝗱𝗲𝗿𝘀. 𝗥𝗲𝗮𝗱 𝗺𝗼𝗿𝗲: 𝗵𝘁𝘁𝗽𝘀://𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆𝗸𝗶𝘁𝗰𝗵𝗲𝗻.𝘀𝘀𝗽𝗻𝗲𝘁.𝗼𝗿𝗴/𝟮𝟬𝟮𝟱/𝟬𝟰/𝟮𝟵/𝗮𝗶-𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆-𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲-𝗮𝗻𝗱-𝗺𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻-𝗶𝗻-𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆-𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴-𝗹𝗲𝘀𝘀𝗼𝗻𝘀-𝗳𝗿𝗼𝗺-𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝘆-𝗳𝗿𝗼𝗻𝘁-𝗿𝘂𝗻𝗻𝗲𝗿𝘀/ #𝗔𝗜 #𝗦𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴
With a higher setting, temperature = 1, I get a more creative result:
🔥 𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗲𝗿𝘀 𝗮𝗿𝗲 𝗺𝗼𝘃𝗶𝗻𝗴 𝗯𝗲𝘆𝗼𝗻𝗱 𝗔𝗜 𝗵𝘆𝗽𝗲—𝘁𝗼𝘄𝗮𝗿𝗱 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆, 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 & 𝗿𝗲𝗮𝗹 𝗥𝗢𝗜. 𝗧𝗵𝗶𝘀 𝗯𝗼𝗹𝗱 𝗦𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆 𝗞𝗶𝘁𝗰𝗵𝗲𝗻 𝗽𝗶𝗲𝗰𝗲 𝗯𝗿𝗲𝗮𝗸𝘀 𝗱𝗼𝘄𝗻 𝘄𝗵𝗮𝘁 𝘁𝗵𝗲 𝗳𝗿𝗼𝗻𝘁-𝗿𝘂𝗻𝗻𝗲𝗿𝘀 𝗮𝗿𝗲 𝗱𝗼𝗶𝗻𝗴 𝗿𝗶𝗴𝗵𝘁 𝗻𝗼𝘄 𝘄𝗶𝘁𝗵 𝗚𝗲𝗻𝗔𝗜. 👀 𝗚𝗲𝘁 𝗶𝗻𝘀𝗽𝗶𝗿𝗲𝗱: 𝗵𝘁𝘁𝗽𝘀://𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆𝗸𝗶𝘁𝗰𝗵𝗲𝗻.𝘀𝘀𝗽𝗻𝗲𝘁.𝗼𝗿𝗴/𝟮𝟬𝟮𝟱/𝟬𝟰/𝟮𝟵/𝗮𝗶-𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆-𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲-𝗮𝗻𝗱-𝗺𝗼𝗻𝗲𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻-𝗶𝗻-𝘀𝗰𝗵𝗼𝗹𝗮𝗿𝗹𝘆-𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴-𝗹𝗲𝘀𝘀𝗼𝗻𝘀-𝗳𝗿𝗼𝗺-𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝘆-𝗳𝗿𝗼𝗻𝘁-𝗿𝘂𝗻𝗻𝗲𝗿𝘀/ #𝗔𝗜𝗶𝗻𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 #𝗚𝗲𝗻𝗔𝗜 #𝗙𝘂𝘁𝘂𝗿𝗲𝗢𝗳𝗣𝘂𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴
[Edited to add: You can set the temperature higher than 1. In the OpenAI API, the maximum is 2. Here’s what the result with temperature = 2:
📚🤖 AI is rewriting the rules of scholarly publishing—literally. From governance to monetization, this Scholarly Kitchen piece dives into how the frontrunners are actually using GenAI. Skip the buzz. Get the blueprint.👉https://scholarlykitchen.sspnet.org/2025/04/29/ai-strategy-governance-and-monetization-in-scholarly-publishing-lessons-from-industry-front-runners/#GenAI #AIstrategy #ScholarlyPublishing
If you’re using the ChatGPT interface, there aren’t strict constraints on what you type—you could even enter 100. The model will probably interpret this as a request for maximum creativity!]
This technique doesn’t always work consistently via the chat interface, but it can be useful when you’re aiming to tailor the tone or creativity level.
